Use this URL to cite or link to this record in EThOS:
Title: Advanced automated PET image segmentation in radiation therapy
Author: Parkinson, Craig
ISNI:       0000 0004 7968 6999
Awarding Body: Cardiff University
Current Institution: Cardiff University
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Manual segmentation of the metabolic tumour volume (MTV) in positron emission tomography (PET) imaging is subject to intra and inter-observer variability. Many PET based automatic segmentation algorithms (PETAS) have been proposed as solutions to this problem with machine-learned techniques showing promise for accurate MTV segmentation. However, no consensus has been reached on the optimal method for radiotherapy (RT) treatment planning, with the current American Association for Physcists in Medicine Task Group 211 and the International Atomic Energy Association advisory committees recommending that not one single PET-AS can be recommended for target volume delineation. This project, therefore, aimed to improve the MTV segmentation of a machine-learned PET-AS methodology called ATLAAS which has been proposed for standardised MTV segmentation by Berthon et al in Radiother Onc (2016). Berthon et al additionally validated the ATLAAS algorithm on diagnostic PET imaging in Radiother Onc (2017). However, it has not been validated externally or for the role of MTV segmentation during treatment. Intratreatment segmentation is challenging due to reduced metabolic uptake, tuiv mour to background ratio and reduced metabolic volume. Therefore, in this body of work, the performance of ATLAAS and 151 other PET-AS chosen from the literature, were evaluated for suitable MTV segmentation in PET imaging acquired after one cycle of neoadjuvant chemotherapy. This research resulted in the development of a new training dataset and demonstrated that ATLAAS can be used as a basis for adaptive radiotherapy and trained on imaging datasets outside of the original training cohort. However, this research still demonstrated that the performance of ATLAAS could be improved. Therefore, this led to an investigation into the inclusion of additional tumour characteristics in the development of the ATLAAS training model, in order to reduce the impact PET image resolution has on MTV segmentation. In this research, derived MTVs were compared to \ground truth" volumes derived from CT imaging. The results presented in this body of work, showed that interpolating PET imaging to the resolution of the CT image improved the performance of PET-AS segmentation and improved ATLAAS MTV segmentation by 19% and inclusion of one of the tumour features compactness one, compactness two or sphericity in the ATLAAS training model improved MTV segmentation by an additional 3%. As part of this body of work, the requirement for a standardised PET-AS method was demonstrated by developing prognostic models, using standardised imaging and tumour features, from the MTV derived by 9 PET-AS demonstrated by Berthon et al in Phys. Med. Biol (2017) to be promising for accurate MTV segmentation. This showed how segmentation of the MTV 120-80% Threshold in increments of 10%, Adaptive Thresholding, Region Growing, K-means Clustering with 2 and 3 clusters, Gaussian Fuzzy C-means with 3 and 4 clusters and Fuzzy-C means with 2 clusters v has a subsequent effect on patient risk stratification with patients changing risk stratification quartiles dependent upon the PET-AS used to derive the MTV.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available